Optimization of neural network structure using involution operator based on particle swarm optimization for image classification

Deep neural networks have made signifi-cant progress in image classification in recent years, however good deep neural networks take a lot of hu-man labor and computational resources, and they must be developed by person with professional expe-rience. Most good deep neural networks now employ convolution operators for feature extraction, however due to convolution spatially agnostic and channel-specific, they lose their capacity to deal with diverse spaces and visual modes. As a result, this article uses a new operator involution based on the inverse con-volution operator's design principle, which is com-bined with the particle swarm optimization algorithm's (PSO) high precision and quick convergence features, as well as the variable length encoding approach. Convolution operator problems can be solved, and the most effective deep neural network structure for the image classification problem can be generated automatically. Experiments demonstrate that the neu-ral network structure created by the method presented in this study outperforms several similar algorithms in terms of recognition accuracy and number of pa-rameters generated, as well as saving a lot of time and computer resources.